Ah, right, the cases where you and your opponent's interests are not perfectly anti-aligned make things a bit trickier, possibly introducing some game theory into the mix. Then I don't know. :)

My first instinct is to say that in principle you should provide the neural net both "must-win" parameters, and have the neural net produce two value outputs, namely the expected utilities for each side separately (which might not sum to 0), which the MCTS would accumulate separately, and at each node depending on who is to move it would use the appropriate side's statistics to choose the child to simulate. That's quite a big difference though, and I haven't thought about ways that this could go wrong, it seems like there might easily be some big pitfalls here. The case where you and your opponent's interests are exactly anti-aligned should still be straightforward though. In that case the way I think of it is that "Play chess where draws are worth half a win and half a loss" and "Play chess where draws are losing for you and winning for your opponent" are two entirely distinct zero-sum games that merely happen to share a lot of rules and features. So of course you should train on both games and distinguish the two so that the neural net always knows which one it's playing, but you can still share the same neural net instead of having two separate nets to take advantage of the fact that each one will regularize the learning for the other. Maybe you still do need to take a little care, for example in Chess if the bot gets sufficiently strong then must-win as black might just always fail to succeed and only produce uninformative samples of always failing, harming the training. I'm optimistic, but ultimately, all this would still need testing. On Tue, Feb 13, 2018 at 12:11 PM, Dan Schmidt <d...@dfan.org> wrote: > Do you intend to use the same draw values for both sides in the self-play > games? They can be independent: > - in a 3/1/0 scenario, neither player is especially happy with a draw > (and in fact would rather each throw a game to each other in a two-game > match than make two draws, but that's a separate issue); > - in a match with one game left, both players agree that a draw and a > Black win (say) are equivalent results; > - in a tournament, the must-win situations of both players could be > independent. > > In real life you usually have a good sense of how your opponent's > "must-win" parameter is set, but that doesn't really apply here. > > > On Tue, Feb 13, 2018 at 10:58 AM, David Wu <lightvec...@gmail.com> wrote: > >> Actually this pretty much solves the whole issue right? Of course the >> proof would be to actually test it out, but it seems to me a pretty >> straightforward solution, not nontrivial at all. >> >> >> On Feb 13, 2018 10:52 AM, "David Wu" <lightvec...@gmail.com> wrote: >> >> Seems to me like you could fix that in the policy too by providing an >> input feature plane that indicates the value of a draw, whether 0 as >> normal, or -1 for must-win, or -1/3 for 3/1/0, or 1 for only-need-not-lose, >> etc. >> >> Then just play games with a variety of values for this parameter in your >> self-play training pipeline so the policy net gets exposed to each kind of >> game. >> >> On Feb 13, 2018 10:40 AM, "Dan Schmidt" <d...@dfan.org> wrote: >> >> The AlphaZero paper says that they just assign values 1, 0, and -1 to >> wins, draws, and losses respectively. This is fine for maximizing your >> expected value over an infinite number of games given the way that chess >> tournaments (to pick the example that I'm familiar with) are typically >> scored, where you get 1, 0.5, and 0 points respectively for wins, draws, >> and losses. >> >> However 1) not all tournaments use this scoring system (3/1/0 is popular >> these days, to discourage draws), and 2) this system doesn't account for >> must-win situations where a draw is as bad as a loss (say you are 1 point >> behind your opponent and it's the last game of a match). Ideally you'd keep >> track of all three probabilities and use some linear meta-scoring function >> on top of them. I don't think it's trivial to extend the AlphaZero >> architecture to handle this, though. Maybe it is sufficient to train with >> the standard meta-scoring (while keeping track of the separate W/D/L >> probabilities) but then use the currently applicable meta-scoring while >> playing. Your policy network won't quite match your current situation, but >> at least your value network and search will. >> >> On Tue, Feb 13, 2018 at 10:05 AM, "Ingo AlthÃ¶fer" <3-hirn-ver...@gmx.de> >> wrote: >> >>> Hello, >>> >>> what is known about proper MCTS procedures for games >>> which do not only have wins and losses, but also draws >>> (like chess, Shogi or Go with integral komi)? >>> >>> Should neural nets provide (win, draw, loss)-probabilities >>> for positions in such games? >>> >>> Ingo. >>> _______________________________________________ >>> Computer-go mailing list >>> Computer-go@computer-go.org >>> http://computer-go.org/mailman/listinfo/computer-go >> >> >> >> _______________________________________________ >> Computer-go mailing list >> Computer-go@computer-go.org >> http://computer-go.org/mailman/listinfo/computer-go >> >> >> >> >> _______________________________________________ >> Computer-go mailing list >> Computer-go@computer-go.org >> http://computer-go.org/mailman/listinfo/computer-go >> > > > _______________________________________________ > Computer-go mailing list > Computer-go@computer-go.org > http://computer-go.org/mailman/listinfo/computer-go >

_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go